def __init__(self, state_size, action_size, n_agents, seed):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
self.stacked_state_size = state_size * n_agents
self.stacked_action_size = action_size * n_agents
# Actor networks
self.actor_local = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=ACTOR_LR)
# Critic networks
self.critic_local = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_target = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=CRITIC_LR)
# OUNoise
self.exploration_noise = OUNoise(action_size, seed)
def __init__(self, state_size, action_size, agent_id):
self.state_size = state_size
self.action_size = action_size
self.seed = args['seed']
self.device = args['device']
#self.args = args
# Q-Network
self.actor_network = ActorNetwork(state_size, action_size).to(self.device)
self.actor_target = ActorNetwork(state_size, action_size).to(self.device)
self.actor_optimizer = optim.Adam(self.actor_network.parameters(), lr=args['LR_ACTOR'])
#Model takes too long to run --> load model weights from previous run (took > 24hours on my machine)
#if not agent_id:
# self.actor_network.load_state_dict(torch.load(args['agent_p0_path']), strict=False)
# self.actor_target.load_state_dict(torch.load(args['agent_p0_path']), strict=False)
#else:
# self.actor_network.load_state_dict(torch.load(args['agent_p1_path']), strict=False)
# self.actor_target.load_state_dict(torch.load(args['agent_p1_path']), strict=False)
# Replay memory
self.memory = ReplayBuffer(action_size, args['BUFFER_SIZE'], args['BATCH_SIZE'], self.device, self.seed)
# Noise process
self.noise = OUNoise(action_size, self.seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
self.mCriticLoss = 0
self.actorLoss = 0
def __init__(self, state_size, obs_size, action_size, num_agents):
super(DDPGAgent, self).__init__()
#self.actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
#self.target_actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.target_critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
self.actor = ActorNetwork(obs_size, action_size).to(device)
self.critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
self.target_actor = ActorNetwork(obs_size, action_size).to(device)
self.target_critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
#self.noise = OUNoise(out_actor, scale=1.0 )
self.noise = OUNoise(action_size, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
self.actor_optimizer = Adam(self.actor.parameters(), lr=LR_ACTOR)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
def __init__(self, state_size, action_size, random_seed, agent_size=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.agent_size = agent_size
self.local_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.target_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.local_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.opt_actor = optim.Adam(self.local_actor.parameters(), lr=LR_ACTOR)
self.opt_critic = optim.Adam(self.local_critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
class Actor():
def __init__(self, state_size, action_size, random_seed, learning_rate,
noise, device):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.learning_rate = learning_rate
self.actor_local = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
random_seed).to(device)
hard_update(self.actor_target, self.actor_local)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=self.learning_rate)
self.noise = noise
self.device = device
def act(self, state, noise_factor, add_noise):
""" Returns actions for given state as per given policy"""
state = torch.from_numpy(state).float().to(self.device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
action += (noise_factor * self.noise.sample())
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# create the local and target actor networks
self.actor_local = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
seed).to(device)
# create the local and target critic networks
self.critic_local = CriticNetwork(state_size, action_size,
seed).to(device)
self.critic_target = CriticNetwork(state_size, action_size,
seed).to(device)
# optimizers for local actor and critic
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR,
weight_decay=0.0)
# MSE loss for updating the critic
self.critic_loss_function = nn.MSELoss()
# ensure that the local and target networks are initialized with the same random weights
for target_param, param in zip(self.actor_target.parameters(),
self.actor_local.parameters()):
target_param.data.copy_(param.data)
for target_param, param in zip(self.critic_target.parameters(),
self.critic_local.parameters()):
target_param.data.copy_(param.data)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
# init the noise class to sample from
self.noise = GaussianNoise(self.action_size)
def create_model(context):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
context.sess = tf.InteractiveSession(config=config)
context.actor = ActorNetwork(
context.sess, [len(context.assets), context.n,
len(context.features)],
len(context.assets) + 1, context.actor_learning_rate, context.tau,
context.minibatch_size)
context.critic = CriticNetwork(
context.sess, [len(context.assets), context.n,
len(context.features)],
len(context.assets) + 1, context.critic_learning_rate, context.tau,
context.gamma, context.actor.get_num_trainable_vars())
# Inicializar las variables de Tensorflow
context.sess.run(tf.global_variables_initializer())
context.saver = tf.train.Saver()
# Inicializar los pesos de las redes objetivo
context.actor.update_target_network()
context.critic.update_target_network()
class Learner:
def __init__(self, opt, q_batch):
self.opt = opt
self.q_batch = q_batch
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
self.actor.share_memory()
self.critic.share_memory()
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=opt.lr)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=opt.lr)
def learning(self):
torch.manual_seed(self.opt.seed)
while True:
# batch-trace
states, actions, rewards = self.q_batch.get(block=True)
onehot_actions = torch.FloatTensor(
index2onehot(actions, self.n_act)).to(self.device)
# update actor network
self.actor_optimizer.zero_grad()
action_log_probs = self.actor(states)
action_log_probs = torch.sum(action_log_probs * onehot_actions, 1)
values = self.critic(states)
advantages = rewards - values.detach()
pg_loss = -torch.sum(action_log_probs * advantages)
actor_loss = pg_loss
actor_loss.backward()
self.actor_optimizer.step()
# update critic network
self.critic_optimizer.zero_grad()
target_values = rewards
critic_loss = nn.MSELoss()(values, target_values)
critic_loss.backward()
self.critic_optimizer.step()
def __init__(self, state_size, action_size, random_seed, learning_rate,
noise, device):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.learning_rate = learning_rate
self.actor_local = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
random_seed).to(device)
hard_update(self.actor_target, self.actor_local)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=self.learning_rate)
self.noise = noise
self.device = device
def build_actor_critic(sess, env):
w_init = tflearn.initializations.xavier_initializer()
with tf.variable_scope("model", reuse=None, initializer=w_init):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=True)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=True)
sess.run(tf.global_variables_initializer())
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
return actor, critic
def __init__(self, opt, q_batch):
self.opt = opt
self.q_batch = q_batch
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
self.actor.share_memory()
self.critic.share_memory()
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=opt.lr)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=opt.lr)
def __init__(self, state_size, action_size, memory, seed=None):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
if seed is not None:
self.seed = seed
# create the local and target actor networks
self.actor_local = ActorNetwork(state_size, action_size, seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size, seed).to(device)
# create the local and target critic networks
self.critic_local = CriticNetwork(state_size, action_size, seed).to(device)
self.critic_target = CriticNetwork(state_size, action_size, seed).to(device)
# optimizers for local actor and critic
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR, weight_decay=0.0)
# MSE loss for updating the critic
# self.critic_loss_function = nn.MSELoss()
self.critic_loss_function = nn.SmoothL1Loss()
# copy the local networks weights to the target network
self.copy_weights_from_local_to_target()
# Replay memory
self.memory = memory
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
# init the noise class to sample from
self.noise = GaussianNoise(self.action_size)
def __init__(self,
state_size,
action_size,
num_agents,
hidden_in_actor=512,
hidden_out_actor=256,
lr_actor=1e-4,
hidden_in_critic=512,
hidden_out_critic=256,
lr_critic=3e-4,
weight_decay_critic=0,
seed=1,
device='cpu'):
super(DDPGAgent, self).__init__()
self.device = device
# Actor
self.actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.target_actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.actor_optimizer = Adam(self.actor.parameters(), lr=lr_actor)
# Target
self.critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=lr_critic,
weight_decay=weight_decay_critic)
# Noise
self.noise = OUNoise(action_size, seed, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
def build_actor_critic(sess, env):
with tf.variable_scope("model", reuse=None):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=False)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=False)
sess.run(tf.global_variables_initializer())
return actor, critic
class DDPGAgent:
#def __init__(self, in_actor=14, hidden_in_actor=16, hidden_out_actor=8, out_actor=2,
#in_critic=20, hidden_in_critic=32, hidden_out_critic=16,
#lr_actor=1.0e-2, lr_critic=1.0e-2):
def __init__(self, state_size, obs_size, action_size, num_agents):
super(DDPGAgent, self).__init__()
#self.actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
#self.target_actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.target_critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
self.actor = ActorNetwork(obs_size, action_size).to(device)
self.critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
self.target_actor = ActorNetwork(obs_size, action_size).to(device)
self.target_critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
#self.noise = OUNoise(out_actor, scale=1.0 )
self.noise = OUNoise(action_size, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
self.actor_optimizer = Adam(self.actor.parameters(), lr=LR_ACTOR)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
def act(self, obs, noise=0.0):
if type(obs) == np.ndarray:
obs = torch.from_numpy(obs).float().to(device)
#self.actor.eval()
action = self.actor(obs)
action += noise * self.noise.noise()
#self.actor.train()
#return action.cpu().data.numpy()
return action
def target_act(self, obs, noise=0.0):
if type(obs) == np.ndarray:
obs = torch.from_numpy(obs).float().to(device)
#obs = obs.to(device)
#self.target_actor.eval()
#action = self.target_actor(obs) + noise*self.noise.noise()
action = self.target_actor(obs)
action += noise * self.noise.noise()
#self.target_actor.train()
#return action.cpu().data.numpy()
return action
def __init__(self, opt, actor_id, q_trace, learner):
self.opt = opt
self.q_trace = q_trace
self.learner = learner
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed + actor_id)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.n_episodes = 0
self.n_steps = 0
self.gamma = opt.gamma
# epsilon
self.eps_greedy = 0.4 ** (1 + actor_id * 7 / (opt.n_actors - 1)) \
if opt.n_actors > 1 else 0.4
# モデル
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
def build_actor_critic(sess, env, env_eval):
w_init = tflearn.initializations.xavier_initializer()
with tf.variable_scope("model", reuse=None, initializer=w_init):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=True)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=True)
# if config.noise_std:
# actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(env.input_cardin * env.state_cardin),
# sigma=config.noise_std,
# sigma_dec=config.noise_dec)
# else:
# actor_noise = None
with tf.variable_scope("model", reuse=True):
with tf.name_scope("actor"):
actor_eval = ActorNetwork(sess,
env_eval,
config,
is_training=False)
with tf.name_scope("critic"):
critic_eval = CriticNetwork(sess,
env_eval,
config,
is_training=False)
sess.run(tf.global_variables_initializer())
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
return actor, critic, actor_eval, critic_eval #, actor_noise
def __init__(self,
device,
key,
state_size, action_size, random_seed,
memory, noise,
lr, weight_decay,
checkpoint_folder = './Saved_Model/'):
self.DEVICE = device
self.KEY = key
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
# Hyperparameters
self.LR = lr
self.WEIGHT_DECAY = weight_decay
self.CHECKPOINT_FOLDER = checkpoint_folder
# Actor Network (w/ Target Network)
self.local = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.target = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.optimizer = optim.Adam(self.local.parameters(), lr=self.LR)
self.checkpoint_full_name = self.CHECKPOINT_FOLDER + 'checkpoint_actor_' + str(self.KEY) + '.pth'
if os.path.isfile(self.checkpoint_full_name):
self.local.load_state_dict(torch.load(self.checkpoint_full_name))
self.target.load_state_dict(torch.load(self.checkpoint_full_name))
# Replay memory
self.memory = memory
# Noise process
self.noise = noise
def __init__(self, opt, actor_id, q_trace, learner):
self.opt = opt
self.q_trace = q_trace
self.learner = learner
self.env = gym.make(self.opt.env)
self.env_state = self.env.reset()
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
## パラメータ
self.batch_size = opt.batch_size
self.roll_out_n_steps = opt.roll_out_n_steps
self.gamma = opt.gamma
self.eps_greedy = 0.4 ** (1 + actor_id * 7 / (opt.n_actors - 1)) \
if opt.n_actors > 1 else 0.4
self.n_episodes = 0
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device) # ActorNetwork
self.critic = CriticNetwork(self.n_state).to(self.device) # CriticNetwork
def __init__(self, path_to_weights, weights_id):
print("Init ANN")
self.init = False
# aviod TF from allocation all GPU mem
# https://stackoverflow.com/questions/34199233/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
#sess = tf.Session()
K.set_session(sess)
self.path_to_weights = path_to_weights
self.next_weights_id = weights_id
#check if path is valid
if not os.path.isdir(self.path_to_weights):
print("Folder with weights does not exist!")
print(self.path_to_weights)
self.init = False
else:
self.init = True
if self.init:
#create actor network
self.actor = ActorNetwork.ActorNetwork(sess, 5, 2)
#load first weights
print("Loading first weights for ANN from " + str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
try:
self.actor.model.load_weights(
str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
self.next_weights_id = self.next_weights_id + 1
except:
print("Cannot find the weight (.h5) file")
print(str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
self.init = False
class DDPGAgent:
def __init__(self,
state_size,
action_size,
num_agents,
hidden_in_actor=512,
hidden_out_actor=256,
lr_actor=1e-4,
hidden_in_critic=512,
hidden_out_critic=256,
lr_critic=3e-4,
weight_decay_critic=0,
seed=1,
device='cpu'):
super(DDPGAgent, self).__init__()
self.device = device
# Actor
self.actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.target_actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.actor_optimizer = Adam(self.actor.parameters(), lr=lr_actor)
# Target
self.critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=lr_critic,
weight_decay=weight_decay_critic)
# Noise
self.noise = OUNoise(action_size, seed, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
def reset(self):
self.noise.reset()
def act(self, obs, noise_factor=0.0):
if torch.is_tensor(obs):
states = obs
else:
states = torch.from_numpy(obs).float().to(self.device)
self.actor.eval()
with torch.no_grad():
actions = self.actor(states).cpu().data.numpy()
self.actor.train()
actions += noise_factor * self.noise.sample()
return np.clip(actions, -1, 1)
def target_act(self, obs):
if torch.is_tensor(obs):
states = obs
else:
states = torch.from_numpy(obs).float().to(self.device)
self.target_actor.eval()
with torch.no_grad():
actions = self.target_actor(states).cpu().data.numpy()
self.target_actor.train()
return np.clip(actions, -1, 1)
def main(args):
with tf.Session() as sess:
env = gym.make(args['env'])
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
# assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
if args['train']:
if not os.path.exists(args['save_dir']):
os.makedirs(args['save_dir'])
with open(os.path.join(args['save_dir'], 'config.json'), 'w') as f:
json.dump(args, f, indent=2)
train(sess, env, args, actor, critic, actor_noise)
else:
# ddpg = []
# indexes = [e for e in range(400) if e % 10 == 9]
# indexes = [0] + indexes
indexes = [399]
num_test_tasks = 100
buckets = 1
successes = []
directory = args['to_pickle']
for index in indexes:
# times = []
task_success = []
saver = tf.train.Saver()
saver.restore(
sess, "../final_models/multitask/fixed/{0}/model-{1}.ckpt".
format(directory, index))
for _ in range(buckets):
tasks = env.unwrapped.sample_tasks(num_test_tasks)
# tasks = [{'goal': np.array([0., 0.])} for e in range(num_test_tasks)]
success = 0
for task in tasks:
s = env.reset_task(task)
step = 0
d = False
while not d:
# env.render()
action = actor.predict_target(
np.reshape(s, (1, actor.s_dim)))[0]
step += 1
s, r, d, _ = env.step(action)
if r == 1:
success += 1
# times.append(step)
env.close()
task_success.append(success / num_test_tasks)
successes.append(task_success)
# ddpg.append(times)
# out = [successes, ddpg]
env.close()
if not os.path.exists('./pkls'):
os.makedirs('./pkls')
with open('./pkls/{0}.pkl'.format(args['save_dir']), 'wb') as f:
pickle.dump(successes, f)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, agent_id, args):
self.state_size = state_size
self.action_size = action_size
self.seed = args['seed']
self.device = args['device']
self.args = args
# Q-Network
self.actor_network = ActorNetwork(state_size, action_size,
args).to(self.device)
self.actor_target = ActorNetwork(state_size, action_size,
args).to(self.device)
self.actor_optimizer = optim.Adam(self.actor_network.parameters(),
lr=args['LR_ACTOR'])
#Model takes too long to run --> load model weights from previous run (took > 24hours on my machine)
if not agent_id:
self.actor_network.load_state_dict(torch.load(
args['agent_p0_path']),
strict=False)
self.actor_target.load_state_dict(torch.load(
args['agent_p0_path']),
strict=False)
else:
self.actor_network.load_state_dict(torch.load(
args['agent_p1_path']),
strict=False)
self.actor_target.load_state_dict(torch.load(
args['agent_p1_path']),
strict=False)
# Replay memory
self.memory = ReplayBuffer(action_size, args['BUFFER_SIZE'],
args['BATCH_SIZE'], self.seed)
# Noise process
self.noise = OUNoise(action_size, self.seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
if len(self.memory) > self.args['BATCH_SIZE']:
experiences = self.memory.sample()
self.train(experiences)
def act(self, current_state):
with torch.no_grad():
self.actor_network.eval()
input_state = torch.from_numpy(current_state).float().to(
self.device)
with torch.no_grad():
action = self.actor_network(input_state).cpu().data.numpy()
self.actor_network.train()
action += self.noise.sample()
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def train(self, experiences):
global states_
global next_states_
global actions_
global max_min_actions_vector
global max_min_states_vector
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
with torch.no_grad():
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next = mCritic.target(next_states, actions_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (GAMMA * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = mCritic.network(states, actions)
mCritic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
mCritic.optimizer.zero_grad()
mCritic_loss.backward()
mCritic.optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_network(states)
actor_loss = -mCritic.network(states, actions_pred).mean()
# Minimize the loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(mCritic.network, mCritic.target, TAU)
self.soft_update(self.actor_network, self.actor_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class Learner(object):
def __init__(self, opt, q_batch):
self.opt = opt
self.q_batch = q_batch
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
self.actor.share_memory()
self.critic.share_memory()
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=opt.lr)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=opt.lr)
def learning(self):
torch.manual_seed(self.opt.seed)
coef_hat = torch.FloatTensor([self.opt.coef_hat]*self.opt.batch_size*self.opt.n_step).view(self.opt.batch_size, self.opt.n_step)
rho_hat = torch.FloatTensor([self.opt.rho_hat]*self.opt.batch_size*self.opt.n_step).view(self.opt.batch_size, self.opt.n_step)
while True:
# batch-trace
states, actions, rewards, dones, action_log_probs = self.q_batch.get(block=True)
logit_log_probs = self.actor(states)
V = self.critic(states).view(self.opt.batch_size, self.opt.n_step) * (1 - dones)
action_probs = torch.exp(action_log_probs)
logit_probs = torch.exp(logit_log_probs)
is_rate = torch.prod(logit_probs / (action_probs + 1e-6), dim=-1).detach()
coef = torch.min(coef_hat, is_rate) * (1 - dones)
rho = torch.min(rho_hat, is_rate) * (1 - dones)
# V-trace
v_trace = torch.zeros((self.opt.batch_size, self.opt.n_step)).to(self.device)
target_V = V.detach()
for rev_step in reversed(range(states.size(1) - 1)):
v_trace[:, rev_step] = target_V[:, rev_step] \
+ rho[:, rev_step] * (rewards[:, rev_step] + self.opt.gamma*target_V[:, rev_step+1] - target_V[:, rev_step]) \
+ self.opt.gamma * coef[:, rev_step] * (v_trace[:, rev_step+1] - target_V[:, rev_step+1])
# actor loss
onehot_actions = torch.FloatTensor(
idx2onehot(actions.cpu().numpy(), self.opt.batch_size, self.n_act)).to(self.device)
logit_log_probs = torch.sum(logit_log_probs * onehot_actions, dim=-1)
advantages = rewards + self.opt.gamma * v_trace - V
pg_loss = -torch.sum(logit_log_probs * advantages.detach())
actor_loss = pg_loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# critic
critic_loss = torch.mean((v_trace.detach() - V)**2)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
class Actor(object):
def __init__(self, opt, actor_id, q_trace, learner):
self.opt = opt
self.q_trace = q_trace
self.learner = learner
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed + actor_id)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.n_episodes = 0
self.n_steps = 0
self.gamma = opt.gamma
# epsilon
self.eps_greedy = 0.4 ** (1 + actor_id * 7 / (opt.n_actors - 1)) \
if opt.n_actors > 1 else 0.4
# モデル
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
def performing(self):
torch.manual_seed(self.opt.seed)
while True:
self.load_model()
self.train_episode()
if self.n_episodes % 100 == 0:
rewards = self.evaluation(self.env)
rewards_mu = np.array(
[np.sum(np.array(l_i), 0) for l_i in rewards]).mean()
print("Episode %d, Average Reward %.2f" %
(self.n_episodes, rewards_mu))
def _softmax_action(self, state):
state = torch.FloatTensor([state]).to(self.device)
softmax_action = torch.exp(self.actor(state)) # expをかけて，行動確率とする
softmax_action = softmax_action.cpu().detach().numpy()
return softmax_action
def exploration_action(self, state):
softmax_action = self._softmax_action(state)
if np.random.rand() > self.eps_greedy:
return np.argmax(softmax_action)
else:
return np.random.choice(self.n_act)
def train_episode(self):
done = False
state = self.env.reset()
self.env_state = state
self.next_done = done
while not done:
self.n_steps += 1
states = np.zeros((self.opt.n_step, self.n_state))
actions = np.zeros(self.opt.n_step)
rewards = np.zeros(self.opt.n_step)
log_probs = np.zeros((self.opt.n_step, self.n_act))
dones = np.ones(self.opt.n_step)
for i in range(self.opt.n_step):
states[i] = self.env_state
dones[i] = self.next_done
log_prob = self.actor(
torch.FloatTensor([state]).to(
self.device)).detach().cpu().numpy()[0]
action = self.exploration_action(state)
next_state, reward, done, info = self.env.step(action)
reward = 0
if done:
if self.n_steps > 190:
reward = 1
else:
reward = -1
log_probs[i] = log_prob
actions[i] = action
rewards[i] = reward
self.env_state = next_state
self.next_done = done
if done:
self.env_state = self.env.reset()
break
# n_step回終了
if done:
self.n_steps = 0
self.n_episodes += 1
self.episode_done = True
else:
self.episode_done = False
self.q_trace.put((states, actions, rewards, dones, log_probs),
block=True)
# choose an action based on state for execution
def action(self, state):
softmax_action = self._softmax_action(state)
action = np.argmax(softmax_action)
return action
def value(self, state): # Qを出力
state_var = torch.FloatTensor([state]).to(self.device)
q_var = self.critic(state_var) # 行動価値を出value
q = q_var.cpu().detach().numpy()
return q
def _discount_reward(self, rewards, final_value):
discounted_r = np.zeros_like(rewards)
R = final_value # Q(s_t, a_t)
for t in reversed(range(0, len(rewards))):
R = rewards[t] + self.gamma * R
discounted_r[t] = R
return discounted_r
def evaluation(self, env_eval):
rewards = []
for i in range(10):
rewards_i = []
state = env_eval.reset()
action = self.action(state)
state, reward, done, _ = env_eval.step(action)
rewards_i.append(reward)
while not done:
action = self.action(state)
state, reward, done, _ = env_eval.step(action)
rewards_i.append(reward)
rewards.append(rewards_i)
return rewards
def load_model(self):
try:
self.actor.load_state_dict(self.learner.actor.state_dict())
self.critic.load_state_dict(self.learner.critic.state_dict())
except:
print('load error')
import torch
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Categorical
from torch.autograd import Variable
import gym
import numpy as np
import matplotlib.pyplot as plt
from model import ActorNetwork, CriticNetwork
actor = ActorNetwork(4, 2)
critic = CriticNetwork(4)
actor_optimizer = optim.Adam(actor.parameters(), lr=1e-4)
critic_optimizer = optim.Adam(critic.parameters(), lr=8e-4)
env = gym.make('CartPole-v0')
GAMMA = 0.99
N_EPISODES = 20000
LOG_STEPS = 100
SAVE_STEPS = 100
def select_action(S):
'''
select action based on currentr state
args:
S: current state
returns:
action to take, log probability of the chosen action
'''
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed, agent_size=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.agent_size = agent_size
self.local_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.target_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.local_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.opt_actor = optim.Adam(self.local_actor.parameters(), lr=LR_ACTOR)
self.opt_critic = optim.Adam(self.local_critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def save_experience(self, state, action, reward, next_state, done):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience
self.memory.add(state, action, reward, next_state, done)
def multi_step(self, t):
# Learn, if enough samples are available in memory
if len(self.memory) > BATCH_SIZE:
if t % 20 == 0:
for i in range(0, 10):
self.learn(self.memory.sample(), GAMMA)
else:
pass
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.local_actor.eval()
with torch.no_grad():
action = self.local_actor(state).cpu().data.numpy()
self.local_actor.train()
if add_noise:
for a in range(0, self.agent_size):
action[a] += self.noise.sample()
return np.clip(action, -1, 1) # all actions between -1 and 1
def reset(self):
self.noise.reset()
def learn(self, experiences, gamma):
"""
Target and Local Critics-Actors are used to sove the moving targets problem.
TargetActor generates the next action, and TargetCritic generates the corresponding Q-value.
This function updates policy and value parameters using given batch of experience tuples.
Q_targets = r + gamma * critic_t(next_state, actor_t(next_state))
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
actions_next = self.target_actor(next_states)
Q_targets_next = self.target_critic(next_states, actions_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.local_critic(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.opt_critic.zero_grad()
critic_loss.backward()
#use gradient clipping when training the critic network
torch.nn.utils.clip_grad_norm_(self.local_critic.parameters(), 1)
self.opt_critic.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.local_actor(states)
actor_loss = -self.local_critic(states, actions_pred).mean()
# Minimize the loss
self.opt_actor.zero_grad()
torch.nn.utils.clip_grad_norm_(self.local_actor.parameters(), 1)
actor_loss.backward()
self.opt_actor.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.local_critic, self.target_critic, TAU)
self.soft_update(self.local_actor, self.target_actor, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
this function manages the update of local and target models syncing
theta_target = tau*theta_local + (1 - tau)*theta_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, memory, seed=None):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
if seed is not None:
self.seed = seed
# create the local and target actor networks
self.actor_local = ActorNetwork(state_size, action_size, seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size, seed).to(device)
# create the local and target critic networks
self.critic_local = CriticNetwork(state_size, action_size, seed).to(device)
self.critic_target = CriticNetwork(state_size, action_size, seed).to(device)
# optimizers for local actor and critic
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR, weight_decay=0.0)
# MSE loss for updating the critic
# self.critic_loss_function = nn.MSELoss()
self.critic_loss_function = nn.SmoothL1Loss()
# copy the local networks weights to the target network
self.copy_weights_from_local_to_target()
# Replay memory
self.memory = memory
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
# init the noise class to sample from
self.noise = GaussianNoise(self.action_size)
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
for _ in range(LEARN_TIMES):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
self.soft_update_all()
def copy_weights_from_local_to_target(self):
# ensure that the local and target networks are initialized with the same random weights
# or copy you saved weights after loading into local
for target_param, param in zip(self.actor_target.parameters(), self.actor_local.parameters()):
target_param.data.copy_(param.data)
for target_param, param in zip(self.critic_target.parameters(), self.critic_local.parameters()):
target_param.data.copy_(param.data)
def act(self, state, add_noise=False):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
# get predicted actions for current state from actor network
self.actor_local.eval()
with torch.no_grad():
action_values = self.actor_local(state)
self.actor_local.train()
# take the predicted actions and add noise, used as exploration in a continuous environment
action_values = action_values.cpu().data.numpy()
if add_noise == True:
action_values += self.noise.sample()
return action_values
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
# unpack the experiences tuple
states, actions, rewards, next_states, dones = experiences
# computer the loss for the actor network per the DDPG algorithm
actor_local_predicted_actions = self.actor_local(states)
policy_loss = -self.critic_local(states, actor_local_predicted_actions).mean()
# compute the loss for the critic network per the DDPG algorithm
predicted_Q_vals = self.critic_local(states, actions)
predicted_actions = self.actor_target(next_states)
Q_next = self.critic_target(next_states, predicted_actions)
Q_targets = rewards + (gamma * Q_next * (1 - dones))
critic_loss = self.critic_loss_function(predicted_Q_vals, Q_targets)
# update the networks
self.actor_optimizer.zero_grad()
policy_loss.backward()
self.actor_optimizer.step()
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
def soft_update_all(self):
# and soft update the target networks
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter # use percent tau local_param.data and rest target_param.data
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau*local_param.data + (1.0-tau)*target_param.data)
class Agent():
def __init__(self, state_size, action_size, n_agents, seed):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
self.stacked_state_size = state_size * n_agents
self.stacked_action_size = action_size * n_agents
# Actor networks
self.actor_local = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=ACTOR_LR)
# Critic networks
self.critic_local = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_target = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=CRITIC_LR)
# OUNoise
self.exploration_noise = OUNoise(action_size, seed)
def act(self, state):
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
# Add exploration noise
action += self.exploration_noise.sample()
return np.clip(action, -1, 1)
def update(self, states, current_agent_states, actions,
current_agent_actions, target_next_actions, rewards,
current_agent_rewards, next_states, dones, current_agent_dones,
action_preds):
flatten_states = torch.reshape(states, shape=(BATCH_SIZE, -1))
flatten_next_states = torch.reshape(next_states,
shape=(BATCH_SIZE, -1))
flatten_actions = torch.reshape(actions, shape=(BATCH_SIZE, -1))
y = current_agent_rewards + GAMMA * self.critic_target(
flatten_next_states,
target_next_actions) * (1 - current_agent_dones)
# Critic loss
critic_loss = F.mse_loss(
y, self.critic_local(flatten_states, flatten_actions))
# Critic backprop
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# Actor loss
actor_loss = -self.critic_local(flatten_states, action_preds).mean()
# Actor backprop
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# Soft updates
self.update_target_network()
def update_target_network(self):
for target_param, local_param in zip(self.actor_target.parameters(),
self.actor_local.parameters()):
target_param.data.copy_(TAU * local_param.data +
(1.0 - TAU) * target_param.data)
for target_param, local_param in zip(self.critic_target.parameters(),
self.critic_local.parameters()):
target_param.data.copy_(TAU * local_param.data +
(1.0 - TAU) * target_param.data)
class Actor:
def __init__(self,
device,
key,
state_size, action_size, random_seed,
memory, noise,
lr, weight_decay,
checkpoint_folder = './Saved_Model/'):
self.DEVICE = device
self.KEY = key
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
# Hyperparameters
self.LR = lr
self.WEIGHT_DECAY = weight_decay
self.CHECKPOINT_FOLDER = checkpoint_folder
# Actor Network (w/ Target Network)
self.local = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.target = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.optimizer = optim.Adam(self.local.parameters(), lr=self.LR)
self.checkpoint_full_name = self.CHECKPOINT_FOLDER + 'checkpoint_actor_' + str(self.KEY) + '.pth'
if os.path.isfile(self.checkpoint_full_name):
self.local.load_state_dict(torch.load(self.checkpoint_full_name))
self.target.load_state_dict(torch.load(self.checkpoint_full_name))
# Replay memory
self.memory = memory
# Noise process
self.noise = noise
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(self.DEVICE)
self.local.eval()
with torch.no_grad():
action = self.local(state).cpu().data.numpy()
self.local.train()
if add_noise:
action += self.noise.sample()
return np.clip(action, -1, 1)
def step(self, state, action, reward, next_state, done):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
self.memory.add(state, action, reward, next_state, done)
def reset(self):
self.noise.reset()
def checkpoint(self):
torch.save(self.local.state_dict(), self.checkpoint_full_name)